metadata
license: cc-by-nc-sa-4.0
task_categories:
- image-feature-extraction
tags:
- interpretability
- vision-transformer
- sparse-autoencoders
- mechanistic-interpretability
ViSAE: Neuroscience-Motivated Probing Suite
This dataset is part of the ViSAE toolbox, presented in the paper Inside the Visual Mind: Neuroscience-Motivated Concept Circuits for Interpreting and Steering Vision Transformers (ICML 2026).
ViSAE is a mechanistic interpretability toolbox designed to decompose Vision Transformer (ViT) representations into human-interpretable concepts using Sparse Autoencoders (SAEs).
- GitHub Repository: deep-real/ViSAE
- Project Page: https://tangli0305.github.io/
Dataset Summary
The probing suite consists of:
- Probing Image Set: 64,000 images designed for SAE training, optimized for high concept coverage (20x more efficient than ImageNet).
- Concept Set: A 16,000 visually grounded concept vocabulary for SAE feature interpretation (available in the GitHub repository).
Usage
To use this dataset for training SAEs or tracing concept circuits, please refer to the official GitHub repository for the implementation details, including:
- Extracting intermediate representations from ViTs.
- Training SAEs for each ViT layer.
- Mapping learned features to concepts using CLIP.
Citation
@inproceedings{li2026visae,
title={Inside the Visual Mind: Neuroscience-Motivated Concept Circuits for Interpreting and Steering Vision Transformers},
author={Li, Tang and Chen, Yanlin and Ma, Mengmeng and Peng, Xi},
booktitle={Proceedings of the International Conference on Machine Learning (ICML)},
year={2026}
}